Skip to content

release: 0.3.0 - closer to the artist's file in every condition, much closer on noisy input - #30

Merged
Stefatorus merged 1 commit into
mainfrom
ccr-73a29992-0f58d2
Oct 11, 2026
Merged

Stefatorus merged 1 commit into
mainfrom
ccr-73a29992-0f58d2

Conversation

@Stefatorus

Copy link
Copy Markdown
Member

What changed

Release 0.3.0. Against v0.2.7 on the regression gate (scorer 4, 246 icons, all eight conditions), every gated axis is better or level.

condition dE00 geom turning gap parameter gap
quality-128ss −12.5 % −8.1 % −24.2 % −24.8 %
quality-512ss −12.1 % −3.3 % −27.8 % −36.5 %
quality-512ssop −15.2 % −3.8 % −4.9 % (level) −35.7 %
quality-web (resized q80 JPEG) −10.4 % −6.4 % −45.4 % −34.3 %
fast-128ss −11.0 % −3.5 % −1.9 % level
fast-512ss −9.1 % −1.5 % −2.5 % level
fast-512ssop −10.3 % −1.8 % level level
fast-web −11.6 % −4.6 % −66.3 % −24.2 %

Quality takes about 1.3× v0.2.7's CPU time (stroke detection on by default, faces completed behind their covers). This was measured on a shared machine. Fast's time is unchanged.

Tracer

  • Sub-pixel boundaries are read from exact column sums.
  • Strokes are on by default in Quality, with merges.
  • Faces are completed behind what covers them. Each completion is accepted only by interval kernels generated from Lean proofs.
  • Counters are cut out of their glyph, and each fill becomes one compound path.
  • Nodes sit at a curve's extrema where the fit prefers them.
  • --editability writes curves the way drawing tools do: arcs as extremum-cut cubics, G1 joins, and fixed handle passes.
  • Noisy input:
    • JPEG halos and codec variants go back to their inks (Fast and Quality).
    • The carve stage leaves the codec band alone.
    • A statistical corner test runs on soft intakes.
    • --restore auto acts only on input that shows damage, and is Quality's default where the model is built in.

Gate and theory

  • New: a geometric match axis, and a noisy JPEG tier (8 conditions).
  • Turning and parameters are now gated as distance from the artist's file.
  • The scorer is faster, with bit-identical numbers.
  • New: a battery of 29 human statistics (bench/human_stats.py).
  • formal/InkvecTheory: Lean proofs. crates/inkvec-verified: checkers generated from them; CI regenerates the checkers and diffs them.
  • New theory pages under docs/theory/.

How verified

  • cargo test --release passes on fit, cli, trace, verified and core (997 tests). The Lean library builds, and the generated checkers regenerate without a diff.
  • cargo fmt --all -- --check is clean. The quality ratchet passes (bench/quality.py --fast), so no new clippy warnings were added.
  • Regression gate: the combined 8-condition run of the final tree passes, and all baselines are re-recorded.
    • A few point estimates moved worse, at most +0.09 % (turning or parameter gap). Each sits beside better or level fidelity, and the project owner accepted them as negligible.
  • Byte identity was used throughout for changes meant to leave output alone (per-merge samples; noted in each commit).
  • docs/THIRD_PARTY.md is current (tools/third_party.py --check).
  • CHANGELOG.md: the [0.3.0] section opens with the numbers above. Every version file is at 0.3.0 (tools/check_versions.py).
  • Contract hashes are re-blessed for x86_64-linux-gnu. The other platforms' hashes come from CI, as for earlier releases.

🤖 Generated with Claude Code

https://claude.ai/code/session_016AZWMqqHhRQ337N5EWhNC6


Generated by Claude Code

…n, much closer on noisy input; faces completed behind their covers; proofs that gate the code

Against v0.2.7 on the regression gate (scorer 4, 246 icons, all eight conditions, every
gated axis better or level): Quality dE00 -12.5 / -12.1 / -15.2 % at 128 / 512 / 512 px
opaque and -10.4 % on the resized quality-80 JPEG tier, with 25-36 % less distance from the
artist's parameter count; Fast dE00 -9.1 to -11.6 %, and on the JPEG tier turning distance
-66 % and parameters 7.29 -> 4.13 per artist parameter. Quality about 1.3x v0.2.7's time;
Fast unchanged.

Sub-pixel boundaries read from exact column sums; strokes on by default with merges; faces
completed behind what covers them (certified by kernels generated from the Lean proofs);
counters cut out of their glyph and one compound path per fill; nodes at a curve's extrema;
--editability writes curves as drawing tools do; noisy input handled (JPEG halos and codec
variants back to their inks, the codec band left alone, --restore auto as Quality's default
where the model is built in, and acting only on damaged input). The gate gains a geometric
match axis, a noisy JPEG tier, distance-from-the-artist axes for turning and parameters, a
faster scorer with the same numbers, and a battery of human statistics.

Verified: 997 Rust tests across fit, cli, trace, verified and core; the Lean library builds
and the generated checkers regenerate without a diff; the combined 8-condition gate passes;
contract hashes re-blessed on Linux; every version file at 0.3.0; third-party list current.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016AZWMqqHhRQ337N5EWhNC6
@Stefatorus
Stefatorus merged commit 3d1841c into main Oct 11, 2026
40 of 42 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants